Palo Alto Networks NetSec-Architect Practice Test 6 - Prepaway
Practice Exams:

Pass Palo Alto Networks NetSec-Architect Exam in First Attempt Guaranteed!

Get 100% Latest Exam Questions, Accurate & Verified Answers!

30 Days Free Updates, Instant Download!

Verified By Experts
Premium File
$69.99 $76.99

NetSec-Architect Premium ETE File


  • Premium File 45 Questions & Answers. Last Update: Nov 10, 2026
Whats Included:
  • Latest Questions
  • 100% Accurate Answers
  • Fast Exam Updates

Palo Alto Networks NetSec-Architect Practice Test 6

Question 50

For fictional Oakbridge Retail Technology, architects investigating ai security design, continuity, policy and architecture governance observe this risk: An AI inference service deploys an externally obtained model image directly into production without model provenance, scanning or approval. Before making broad configuration or topology changes, which finding most directly isolates the failure mechanism?

  1. For Oakbridge Retail Technology and the unsigned foundation model reaches production assessment, Cloud egress and DNS traces show model API requests bypassing the designated AI security inspection point.
  2. For Oakbridge Retail Technology and the unsigned foundation model reaches production assessment, The failover runbook and route tables show secondary inference endpoints use a direct internet route without equivalent controls.
  3. For Oakbridge Retail Technology and the unsigned foundation model reaches production assessment, The approved SaaS classification exists yet the policy has no context-aware Enterprise DLP classifier for that sensitive dataset.
  4. For Oakbridge Retail Technology and the unsigned foundation model reaches production assessment, Model registry audit records no signed artifact digest, vulnerability assessment or predeployment model security finding.

Explanation

Correct Answer

D:

The evidence shows that model registry audit records no signed artifact digest, vulnerability assessment or predeployment model security finding. AI security starts before inference by validating model artifacts and their supply-chain integrity. The direct correction is: Introduce model provenance validation and Prisma AIRS model scanning with approval gates before trusted runtime deployment.

Incorrect Answers

A:

A competing observation indicates that cloud egress and DNS traces show model API requests bypassing the designated AI security inspection point, which points away from this case. The distinguishing evidence is: Model registry audit records no signed artifact digest, vulnerability assessment or predeployment model security finding. That is consistent with the mechanism: AI security starts before inference by validating model artifacts and their supply-chain integrity.

B:

An alternative finding indicates that the failover runbook and route tables show secondary inference endpoints use a direct internet route without equivalent controls; this is not the controlling evidence in the report. Model registry audit records no signed artifact digest, vulnerability assessment or predeployment model security finding. The appropriate correction is: Introduce model provenance validation and Prisma AIRS model scanning with approval gates before trusted runtime deployment.

C:

A different observation—the approved SaaS classification exists yet the policy has no context-aware Enterprise DLP classifier for that sensitive dataset—does not establish the specific fault described here. The record instead shows this: Model registry audit records no signed artifact digest, vulnerability assessment or predeployment model security finding. AI security starts before inference by validating model artifacts and their supply-chain integrity.

Question 51

At fictional Oakbridge Retail Technology, the security design for ai security design, continuity, policy and architecture governance faces: An AI inference service deploys an externally obtained model image directly into production without model provenance, scanning or approval. Diagnostic evidence: Model registry audit records no signed artifact digest, vulnerability assessment or predeployment model security finding. Proposed design change: Introduce model provenance validation and Prisma AIRS model scanning with approval gates before trusted runtime deployment. Which independent observation would demonstrate that the change met its intended outcome?

  1. For Oakbridge Retail Technology and the unsigned foundation model reaches production assessment, A known noncompliant test model is quarantined, while a signed reviewed model is approved and deployed with traceable evidence.
  2. For Oakbridge Retail Technology and the unsigned foundation model reaches production assessment, A forced regional switch preserves authorized inference service and blocks the same prohibited test payload in both regions.
  3. For Oakbridge Retail Technology and the unsigned foundation model reaches production assessment, Benign prompts pass while a synthetic regulated identifier upload is blocked and logged without exposing the actual content in logs.
  4. For Oakbridge Retail Technology and the unsigned foundation model reaches production assessment, Controlled sanctioned model calls are inspected and attributed; unauthorized direct calls are blocked or alerted with evidence.

Explanation

Correct Answer

A:

Verify the result with: A known noncompliant test model is quarantined, while a signed reviewed model is approved and deployed with traceable evidence. This tests the correction against the actual evidence: Model registry audit records no signed artifact digest, vulnerability assessment or predeployment model security finding. AI security starts before inference by validating model artifacts and their supply-chain integrity.

Incorrect Answers

B:

A different check—A forced regional switch preserves authorized inference service and blocks the same prohibited test payload in both regions—would not demonstrate that this particular failure was corrected. Acceptance instead requires this result: A known noncompliant test model is quarantined, while a signed reviewed model is approved and deployed with traceable evidence. This tests the condition shown by model registry audit records no signed artifact digest, vulnerability assessment or predeployment model security finding.

C:

An alternative result would show that benign prompts pass while a synthetic regulated identifier upload is blocked and logged without exposing the actual content in logs; it measures a different outcome. To check this fault, verify: A known noncompliant test model is quarantined, while a signed reviewed model is approved and deployed with traceable evidence. The earlier evidence was that model registry audit records no signed artifact digest, vulnerability assessment or predeployment model security finding.

D:

An alternative check—controlled sanctioned model calls are inspected and attributed; unauthorized direct calls are blocked or alerted with evidence—is not sufficient closure evidence for this fault. The relevant acceptance test is: A known noncompliant test model is quarantined, while a signed reviewed model is approved and deployed with traceable evidence. AI security starts before inference by validating model artifacts and their supply-chain integrity.

Question 52

For the architecture board of fictional Oakbridge Retail Technology, a proposed change to ai security design, continuity, policy and architecture governance follows this problem: An AI inference service deploys an externally obtained model image directly into production without model provenance, scanning or approval. Discriminating observation: Model registry audit records no signed artifact digest, vulnerability assessment or predeployment model security finding. Which network security architecture principle best explains why the existing design is insufficient?

  1. For Oakbridge Retail Technology and the unsigned foundation model reaches production assessment, An AI security architecture needs a known enforcement path for every relevant inference request.
  2. For Oakbridge Retail Technology and the unsigned foundation model reaches production assessment, AI security starts before inference by validating model artifacts and their supply-chain integrity.
  3. For Oakbridge Retail Technology and the unsigned foundation model reaches production assessment, AI runtime security must cover lateral network flows and workloads, not just prompt text at an API endpoint.
  4. For Oakbridge Retail Technology and the unsigned foundation model reaches production assessment, Sanctioning an AI app does not eliminate data loss risk from its specific features and transactions.

Explanation

Correct Answer

B:

AI security starts before inference by validating model artifacts and their supply-chain integrity. That distinction fits the observed clue: model registry audit records no signed artifact digest, vulnerability assessment or predeployment model security finding. The practical fix is to introduce model provenance validation and Prisma AIRS model scanning with approval gates before trusted runtime deployment.

Incorrect Answers

A:

That alternative rule—an AI security architecture needs a known enforcement path for every relevant inference request—does not explain the deciding observation: model registry audit records no signed artifact digest, vulnerability assessment or predeployment model security finding. AI security starts before inference by validating model artifacts and their supply-chain integrity. That is the mechanism the corrective action must address.

C:

The cited mechanism, AI runtime security must cover lateral network flows and workloads, not just prompt text at an API endpoint, does not account for the recorded clue. Model registry audit records no signed artifact digest, vulnerability assessment or predeployment model security finding. AI security starts before inference by validating model artifacts and their supply-chain integrity.

D:

That alternative rule—sanctioning an AI app does not eliminate data loss risk from its specific features and transactions—does not explain the deciding observation: model registry audit records no signed artifact digest, vulnerability assessment or predeployment model security finding. AI security starts before inference by validating model artifacts and their supply-chain integrity. That is the mechanism the corrective action must address.

Question 53

At fictional Eastfield Data Platforms, the architecture team is responsible for cross-region ai security design validation during staged modernization and must verify failure behavior and compliance without expanding unnecessary application exposure. Risk case: An AI inference service deploys an externally obtained model image directly into production without model provenance, scanning or approval. Verified finding: Model registry audit records no signed artifact digest, vulnerability assessment or predeployment model security finding. Which scoped architecture decision most directly addresses the documented control gap without unnecessarily changing other services?

  1. For Eastfield Data Platforms and the unsigned foundation model reaches production assessment, Add sanctioned-app usage controls and a narrowly scoped Enterprise DLP classifier, with training and authorized exception workflows.
  2. For Eastfield Data Platforms and the unsigned foundation model reaches production assessment, Introduce model provenance validation and Prisma AIRS model scanning with approval gates before trusted runtime deployment.
  3. For Eastfield Data Platforms and the unsigned foundation model reaches production assessment, Extend approved regional AI security enforcement, routing and policy synchronization to disaster-recovery inference paths.
  4. For Eastfield Data Platforms and the unsigned foundation model reaches production assessment, Design an appropriately deployed AI security gateway with supported routed/proxy coverage, identity correlation and policy control for approved AI endpoints.

Explanation

Correct Answer

B:

The evidence is that model registry audit records no signed artifact digest, vulnerability assessment or predeployment model security finding. The appropriate change is to introduce model provenance validation and Prisma AIRS model scanning with approval gates before trusted runtime deployment, because AI security starts before inference by validating model artifacts and their supply-chain integrity. After the change, A known noncompliant test model is quarantined, while a signed reviewed model is approved and deployed with traceable evidence.

Incorrect Answers

A:

One alternative is to add sanctioned-app usage controls and a narrowly scoped Enterprise DLP classifier, with training and authorized exception workflows, but this does not address the decisive finding: Model registry audit records no signed artifact digest, vulnerability assessment or predeployment model security finding. Instead, the correction is: Introduce model provenance validation and Prisma AIRS model scanning with approval gates before trusted runtime deployment. Check the result using: A known noncompliant test model is quarantined, while a signed reviewed model is approved and deployed with traceable evidence.

C:

The proposed change—extend approved regional AI security enforcement, routing and policy synchronization to disaster-recovery inference paths—does not repair the condition documented in the case. The decisive evidence is that model registry audit records no signed artifact digest, vulnerability assessment or predeployment model security finding. The necessary change is: Introduce model provenance validation and Prisma AIRS model scanning with approval gates before trusted runtime deployment.

D:

One alternative is to design an appropriately deployed AI security gateway with supported routed/proxy coverage, identity correlation and policy control for approved AI endpoints, but this does not address the decisive finding: Model registry audit records no signed artifact digest, vulnerability assessment or predeployment model security finding. Instead, the correction is: Introduce model provenance validation and Prisma AIRS model scanning with approval gates before trusted runtime deployment. Check the result using: A known noncompliant test model is quarantined, while a signed reviewed model is approved and deployed with traceable evidence.

Question 54

For fictional Eastfield Data Platforms, architects investigating cross-region ai security design validation during staged modernization observe this risk: An AI inference service deploys an externally obtained model image directly into production without model provenance, scanning or approval. Before making broad configuration or topology changes, which finding most directly isolates the failure mechanism?

  1. For Eastfield Data Platforms and the unsigned foundation model reaches production assessment, Model registry audit records no signed artifact digest, vulnerability assessment or predeployment model security finding.
  2. For Eastfield Data Platforms and the unsigned foundation model reaches production assessment, Procurement and processing records lack region, retention and data-flow attributes against the firm's approved governance register.
  3. For Eastfield Data Platforms and the unsigned foundation model reaches production assessment, The approved SaaS classification exists yet the policy has no context-aware Enterprise DLP classifier for that sensitive dataset.
  4. For Eastfield Data Platforms and the unsigned foundation model reaches production assessment, Test transcripts show a prompt-injection string in lower-trust retrieval content preceding a privileged tool request.

Explanation

Correct Answer

A:

The record shows that model registry audit records no signed artifact digest, vulnerability assessment or predeployment model security finding. AI security starts before inference by validating model artifacts and their supply-chain integrity. The corresponding correction is: Introduce model provenance validation and Prisma AIRS model scanning with approval gates before trusted runtime deployment.

Incorrect Answers

B:

A different observation—procurement and processing records lack region, retention and data-flow attributes against the firm's approved governance register—does not establish the specific fault described here. The record instead shows this: Model registry audit records no signed artifact digest, vulnerability assessment or predeployment model security finding. AI security starts before inference by validating model artifacts and their supply-chain integrity.

C:

A competing observation indicates that the approved SaaS classification exists yet the policy has no context-aware Enterprise DLP classifier for that sensitive dataset, which points away from this case. The distinguishing evidence is: Model registry audit records no signed artifact digest, vulnerability assessment or predeployment model security finding. That is consistent with the mechanism: AI security starts before inference by validating model artifacts and their supply-chain integrity.

D:

An alternative finding indicates that test transcripts show a prompt-injection string in lower-trust retrieval content preceding a privileged tool request; this is not the controlling evidence in the report. Model registry audit records no signed artifact digest, vulnerability assessment or predeployment model security finding. The appropriate correction is: Introduce model provenance validation and Prisma AIRS model scanning with approval gates before trusted runtime deployment.

Question 55

At fictional Eastfield Data Platforms, the security design for cross-region ai security design validation during staged modernization faces: An AI inference service deploys an externally obtained model image directly into production without model provenance, scanning or approval. Diagnostic evidence: Model registry audit records no signed artifact digest, vulnerability assessment or predeployment model security finding. Proposed design change: Introduce model provenance validation and Prisma AIRS model scanning with approval gates before trusted runtime deployment. Which independent observation would demonstrate that the change met its intended outcome?

  1. For Eastfield Data Platforms and the unsigned foundation model reaches production assessment, Adversarial test prompts are rejected or contained and authorized customer lookup still works without exporting secrets.
  2. For Eastfield Data Platforms and the unsigned foundation model reaches production assessment, A known noncompliant test model is quarantined, while a signed reviewed model is approved and deployed with traceable evidence.
  3. For Eastfield Data Platforms and the unsigned foundation model reaches production assessment, A forced regional switch preserves authorized inference service and blocks the same prohibited test payload in both regions.
  4. For Eastfield Data Platforms and the unsigned foundation model reaches production assessment, Controlled sanctioned model calls are inspected and attributed; unauthorized direct calls are blocked or alerted with evidence.

Explanation

Correct Answer

B:

A known noncompliant test model is quarantined, while a signed reviewed model is approved and deployed with traceable evidence. This post-change check addresses the recorded finding: model registry audit records no signed artifact digest, vulnerability assessment or predeployment model security finding. AI security starts before inference by validating model artifacts and their supply-chain integrity.

Incorrect Answers

A:

A different check—adversarial test prompts are rejected or contained and authorized customer lookup still works without exporting secrets—would not demonstrate that this particular failure was corrected. Acceptance instead requires this result: A known noncompliant test model is quarantined, while a signed reviewed model is approved and deployed with traceable evidence. This tests the condition shown by model registry audit records no signed artifact digest, vulnerability assessment or predeployment model security finding.

C:

An alternative check—A forced regional switch preserves authorized inference service and blocks the same prohibited test payload in both regions—is not sufficient closure evidence for this fault. The relevant acceptance test is: A known noncompliant test model is quarantined, while a signed reviewed model is approved and deployed with traceable evidence. AI security starts before inference by validating model artifacts and their supply-chain integrity.

D:

A different check—controlled sanctioned model calls are inspected and attributed; unauthorized direct calls are blocked or alerted with evidence—would not demonstrate that this particular failure was corrected. Acceptance instead requires this result: A known noncompliant test model is quarantined, while a signed reviewed model is approved and deployed with traceable evidence. This tests the condition shown by model registry audit records no signed artifact digest, vulnerability assessment or predeployment model security finding.

Question 56

For the architecture board of fictional Eastfield Data Platforms, a proposed change to cross-region ai security design validation during staged modernization follows this problem: An AI inference service deploys an externally obtained model image directly into production without model provenance, scanning or approval. Discriminating observation: Model registry audit records no signed artifact digest, vulnerability assessment or predeployment model security finding. Which network security architecture principle best explains why the existing design is insufficient?

  1. For Eastfield Data Platforms and the unsigned foundation model reaches production assessment, Security design for AI spans the complete active and failover topology, not only the primary region.
  2. For Eastfield Data Platforms and the unsigned foundation model reaches production assessment, AI governance ties technical controls to jurisdictional data handling, consent and audit obligations.
  3. For Eastfield Data Platforms and the unsigned foundation model reaches production assessment, Sanctioning an AI app does not eliminate data loss risk from its specific features and transactions.
  4. For Eastfield Data Platforms and the unsigned foundation model reaches production assessment, AI security starts before inference by validating model artifacts and their supply-chain integrity.

Explanation

Correct Answer

D:

The governing principle is: AI security starts before inference by validating model artifacts and their supply-chain integrity. The case specifically shows model registry audit records no signed artifact digest, vulnerability assessment or predeployment model security finding. The corresponding change is: Introduce model provenance validation and Prisma AIRS model scanning with approval gates before trusted runtime deployment.

Incorrect Answers

A:

An alternative principle holds that security design for AI spans the complete active and failover topology, not only the primary region, but this is not the mechanism shown here. This case reports that model registry audit records no signed artifact digest, vulnerability assessment or predeployment model security finding. Its governing principle is that AI security starts before inference by validating model artifacts and their supply-chain integrity.

B:

The cited mechanism, AI governance ties technical controls to jurisdictional data handling, consent and audit obligations, does not account for the recorded clue. Model registry audit records no signed artifact digest, vulnerability assessment or predeployment model security finding. AI security starts before inference by validating model artifacts and their supply-chain integrity.

C:

At Eastfield Data Platforms, the offered rule—sanctioning an AI app does not eliminate data loss risk from its specific features and transactions—does not explain the deciding observation: model registry audit records no signed artifact digest, vulnerability assessment or predeployment model security finding. AI security starts before inference by validating model artifacts and their supply-chain integrity. That is the mechanism the corrective action must address.

Question 57

At fictional Seabright Global Systems, the architecture team is responsible for ai security design, continuity, policy and architecture governance and must preserve least privilege, auditable change and measurable business availability. Risk case: A support assistant is persuaded by retrieved content to call an unrelated export tool with customer secrets. Verified finding: Test transcripts show a prompt-injection string in lower-trust retrieval content preceding a privileged tool request. Which scoped architecture decision most directly addresses the documented control gap without unnecessarily changing other services?

  1. For Seabright Global Systems and the prompt-injection exfiltration on assistant tool assessment, Extend approved regional AI security enforcement, routing and policy synchronization to disaster-recovery inference paths.
  2. For Seabright Global Systems and the prompt-injection exfiltration on assistant tool assessment, Add sanctioned-app usage controls and a narrowly scoped Enterprise DLP classifier, with training and authorized exception workflows.
  3. For Seabright Global Systems and the prompt-injection exfiltration on assistant tool assessment, Apply workload identity and Kubernetes-aware segmentation with restricted egress paths and AI runtime policy at supported enforcement points.
  4. For Seabright Global Systems and the prompt-injection exfiltration on assistant tool assessment, Architect AI runtime inspection, contextual tool authorization and DLP rules so untrusted retrieved instructions cannot override system-approved capabilities.

Explanation

Correct Answer

D:

Because test transcripts show a prompt-injection string in lower-trust retrieval content preceding a privileged tool request, the relevant step is to architect AI runtime inspection, contextual tool authorization and DLP rules so untrusted retrieved instructions cannot override system-approved capabilities. An AI gateway and per-tool least privilege can prevent model output from becoming unchecked authority. Recovery should be checked using this result: Adversarial test prompts are rejected or contained and authorized customer lookup still works without exporting secrets.

Incorrect Answers

A:

An alternative operation—extend approved regional AI security enforcement, routing and policy synchronization to disaster-recovery inference paths—does not directly resolve the recorded failure. Test transcripts show a prompt-injection string in lower-trust retrieval content preceding a privileged tool request. The appropriate intervention is: Architect AI runtime inspection, contextual tool authorization and DLP rules so untrusted retrieved instructions cannot override system-approved capabilities.

B:

The proposed change—add sanctioned-app usage controls and a narrowly scoped Enterprise DLP classifier, with training and authorized exception workflows—does not repair the condition documented in the case. The decisive evidence is that test transcripts show a prompt-injection string in lower-trust retrieval content preceding a privileged tool request. The necessary change is: Architect AI runtime inspection, contextual tool authorization and DLP rules so untrusted retrieved instructions cannot override system-approved capabilities.

C:

One alternative is to apply workload identity and Kubernetes-aware segmentation with restricted egress paths and AI runtime policy at supported enforcement points, but this does not address the decisive finding: Test transcripts show a prompt-injection string in lower-trust retrieval content preceding a privileged tool request. Instead, the correction is: Architect AI runtime inspection, contextual tool authorization and DLP rules so untrusted retrieved instructions cannot override system-approved capabilities. Check the result using: Adversarial test prompts are rejected or contained and authorized customer lookup still works without exporting secrets.

Question 58

For fictional Seabright Global Systems, architects investigating ai security design, continuity, policy and architecture governance observe this risk: A support assistant is persuaded by retrieved content to call an unrelated export tool with customer secrets. Before making broad configuration or topology changes, which finding most directly isolates the failure mechanism?

  1. For Seabright Global Systems and the prompt-injection exfiltration on assistant tool assessment, Cloud egress and DNS traces show model API requests bypassing the designated AI security inspection point.
  2. For Seabright Global Systems and the prompt-injection exfiltration on assistant tool assessment, Procurement and processing records lack region, retention and data-flow attributes against the firm's approved governance register.
  3. For Seabright Global Systems and the prompt-injection exfiltration on assistant tool assessment, Test transcripts show a prompt-injection string in lower-trust retrieval content preceding a privileged tool request.
  4. For Seabright Global Systems and the prompt-injection exfiltration on assistant tool assessment, The approved SaaS classification exists yet the policy has no context-aware Enterprise DLP classifier for that sensitive dataset.

Explanation

Correct Answer

C:

Test transcripts show a prompt-injection string in lower-trust retrieval content preceding a privileged tool request. This observation matters because an AI gateway and per-tool least privilege can prevent model output from becoming unchecked authority. The corresponding targeted response is to architect AI runtime inspection, contextual tool authorization and DLP rules so untrusted retrieved instructions cannot override system-approved capabilities.

Incorrect Answers

A:

A different observation—cloud egress and DNS traces show model API requests bypassing the designated AI security inspection point—does not establish the specific fault described here. The record instead shows this: Test transcripts show a prompt-injection string in lower-trust retrieval content preceding a privileged tool request. An AI gateway and per-tool least privilege can prevent model output from becoming unchecked authority.

B:

A competing observation indicates that procurement and processing records lack region, retention and data-flow attributes against the firm's approved governance register, which points away from this case. The distinguishing evidence is: Test transcripts show a prompt-injection string in lower-trust retrieval content preceding a privileged tool request. That is consistent with the mechanism: an AI gateway and per-tool least privilege can prevent model output from becoming unchecked authority.

D:

A different observation—the approved SaaS classification exists yet the policy has no context-aware Enterprise DLP classifier for that sensitive dataset—does not establish the specific fault described here. The record instead shows this: Test transcripts show a prompt-injection string in lower-trust retrieval content preceding a privileged tool request. An AI gateway and per-tool least privilege can prevent model output from becoming unchecked authority.

Question 59

At fictional Seabright Global Systems, the security design for ai security design, continuity, policy and architecture governance faces: A support assistant is persuaded by retrieved content to call an unrelated export tool with customer secrets. Diagnostic evidence: Test transcripts show a prompt-injection string in lower-trust retrieval content preceding a privileged tool request. Proposed design change: Architect AI runtime inspection, contextual tool authorization and DLP rules so untrusted retrieved instructions cannot override system-approved capabilities. Which independent observation would demonstrate that the change met its intended outcome?

  1. For Seabright Global Systems and the prompt-injection exfiltration on assistant tool assessment, Required inference dependencies remain functional while simulated cross-namespace regulated-data access is blocked.
  2. For Seabright Global Systems and the prompt-injection exfiltration on assistant tool assessment, Controlled sanctioned model calls are inspected and attributed; unauthorized direct calls are blocked or alerted with evidence.
  3. For Seabright Global Systems and the prompt-injection exfiltration on assistant tool assessment, Adversarial test prompts are rejected or contained and authorized customer lookup still works without exporting secrets.
  4. For Seabright Global Systems and the prompt-injection exfiltration on assistant tool assessment, A forced regional switch preserves authorized inference service and blocks the same prohibited test payload in both regions.

Explanation

Correct Answer

C:

Verify the change by checking: Adversarial test prompts are rejected or contained and authorized customer lookup still works without exporting secrets. The test is tied to the original clue: test transcripts show a prompt-injection string in lower-trust retrieval content preceding a privileged tool request. The correction being tested is: Architect AI runtime inspection, contextual tool authorization and DLP rules so untrusted retrieved instructions cannot override system-approved capabilities.

Incorrect Answers

A:

An alternative result would show that required inference dependencies remain functional while simulated cross-namespace regulated-data access is blocked; it measures a different outcome. To check this fault, verify: Adversarial test prompts are rejected or contained and authorized customer lookup still works without exporting secrets. The earlier evidence was that test transcripts show a prompt-injection string in lower-trust retrieval content preceding a privileged tool request.

B:

An alternative check—controlled sanctioned model calls are inspected and attributed; unauthorized direct calls are blocked or alerted with evidence—is not sufficient closure evidence for this fault. The relevant acceptance test is: Adversarial test prompts are rejected or contained and authorized customer lookup still works without exporting secrets. An AI gateway and per-tool least privilege can prevent model output from becoming unchecked authority.

D:

An alternative result would show that A forced regional switch preserves authorized inference service and blocks the same prohibited test payload in both regions; it measures a different outcome. To check this fault, verify: Adversarial test prompts are rejected or contained and authorized customer lookup still works without exporting secrets. The earlier evidence was that test transcripts show a prompt-injection string in lower-trust retrieval content preceding a privileged tool request.

Question 60

For the architecture board of fictional Seabright Global Systems, a proposed change to ai security design, continuity, policy and architecture governance follows this problem: A support assistant is persuaded by retrieved content to call an unrelated export tool with customer secrets. Discriminating observation: Test transcripts show a prompt-injection string in lower-trust retrieval content preceding a privileged tool request. Which network security architecture principle best explains why the existing design is insufficient?

  1. For Seabright Global Systems and the prompt-injection exfiltration on assistant tool assessment, An AI gateway and per-tool least privilege can prevent model output from becoming unchecked authority.
  2. For Seabright Global Systems and the prompt-injection exfiltration on assistant tool assessment, An AI security architecture needs a known enforcement path for every relevant inference request.
  3. For Seabright Global Systems and the prompt-injection exfiltration on assistant tool assessment, AI runtime security must cover lateral network flows and workloads, not just prompt text at an API endpoint.
  4. For Seabright Global Systems and the prompt-injection exfiltration on assistant tool assessment, AI governance ties technical controls to jurisdictional data handling, consent and audit obligations.

Explanation

Correct Answer

A:

An AI gateway and per-tool least privilege can prevent model output from becoming unchecked authority. In the reported case, test transcripts show a prompt-injection string in lower-trust retrieval content preceding a privileged tool request. The remedy follows that mechanism: Architect AI runtime inspection, contextual tool authorization and DLP rules so untrusted retrieved instructions cannot override system-approved capabilities.

Incorrect Answers

B:

That alternative rule—an AI security architecture needs a known enforcement path for every relevant inference request—does not explain the deciding observation: test transcripts show a prompt-injection string in lower-trust retrieval content preceding a privileged tool request. An AI gateway and per-tool least privilege can prevent model output from becoming unchecked authority. That is the mechanism the corrective action must address.

C:

An alternative principle holds that AI runtime security must cover lateral network flows and workloads, not just prompt text at an API endpoint, but this is not the mechanism shown here. This case reports that test transcripts show a prompt-injection string in lower-trust retrieval content preceding a privileged tool request. Its governing principle is that an AI gateway and per-tool least privilege can prevent model output from becoming unchecked authority.

D:

The cited mechanism, AI governance ties technical controls to jurisdictional data handling, consent and audit obligations, does not account for the recorded clue. Test transcripts show a prompt-injection string in lower-trust retrieval content preceding a privileged tool request. An AI gateway and per-tool least privilege can prevent model output from becoming unchecked authority.

Palo Alto Networks Network Security Architect (NetSec-Architect) Practice Tests